Classification of Cognitive Ability of Healthy Elderly Individuals Using Resting-State Functional Connectivity Magnetic Resonance Imaging and An Extreme Learning Machine

Purpose Quantitative determination of the correlation between cognitive ability and functional biomarkers in the elderly brain is essential. To identify biomarkers associated with cognitive performance in the elderly, this study combined an index model specific for resting-state functional connectivity (FC) magnetic resonance imaging (fcMRI) with a supervised machine learning method. Methods Performance scores on conventional cognitive test batteries and MRI data were obtained for 98 healthy elderly individuals and 90 healthy youth from two public databases. Based on the test scores, the elderly cohort was categorized into two groups: excellent and poor. An fcMRI index model was constructed for each elderly individual to determine the relative differences in FC among brain regions compared with that in the youth cohort. Brain areas sensitive to test scores could then be identified using the fcMRI indexes. To confirm the effectiveness of constructed model, the fcMRI indexes of these brain areas were used as feature matrix inputs for training an extreme learning machine. Classification accuracy was then tested in separate groups and confirmed by N-fold cross-validation. Results This learning study could effectively classify the cognitive status of healthy elderly individuals according to frontal lobe, temporal lobe, and parietal lobe FC values with a mean accuracy of 83.5%, which is substantially higher than that achieved using conventional correlation analysis. Conclusion This fcMRI classification study may facilitate early detection of age-related cognitive decline as well as help reveal the underlying pathological mechanisms.


1． Introduction
The increase in the aging population in many countries has led to an increase in the prevalence of utilizes advanced technology to provide objective measurements of cognitive functions [7,8].
Furthermore, fMRI evaluations can be conducted both during tasks engaging specific neural networks or in the resting state to provide an unbiased (taskindependent) assessment of network activity [9,10].
Resting-state functional connectivity (FC) MRI (fcMRI) can reveal the FC of the whole brain by measuring the spatiotemporal correlations in regional brain activity established by Hebbian-like learning mechanisms [11][12][13][14][15][16]. In addition, fcMRI is noninvasive and objective, and in contrast to neuropsychiatric tests and task-dependent fMRI, can be conducted quickly (in less than 15
(4) The brain extraction tool was used for skull peeling [23]. (5) Nonlinear normalization through continuous rigid body registration was performed. was conducted. When SPM software was used for FC analysis, co-variables in fcMRI data needed to be removed, such as head movement parameters, whole-brain signals, WM signals, cerebrospinal fluid signals, and other relevant co-variables, e.g., age, sex, etc.
All rs-fMRI data were acquired using a 1.5 T Siemens Magnetom Avanto 12 channel head coil scanner and a bold oxygen level-dependent (BOLD) echo planar imaging sequence, yielding 3.5 × 3.5 × 3.5 mm voxels for the elderly group and 3 × 3 × 3 mm voxels for the youth group. Voxels were then assigned to 116 brain regions according to the automatic anatomical labeling (AAL) atlas. The mean BOLD signal of each voxel belonging to a given AAL region was subjected to post-processing analysis [24].

2.3.1． fcMRI
Traditionally, the Pearson correlation coefficient is used to quantify the degree of dependence between regional brain activity patterns (time series) assuming a static BOLD signal and inclusion of all time series for fcMRI estimation [25].
A correlation coefficient of >0 implies that excitation in one area is associated with excitation in another (and vice versa) [24], whereas a coefficient of <0 implies that excitation in one region is associated with inhibition in another. A coefficient of 1 implies self-correlation, whereas a coefficient of 0 implies no correlation between two brain areas.

2.3.2． The fcMRI index model
To identify functional neuroimaging markers with sensitivity to cognitive test scores, a fcMRI index model was constructed from the differences in regional fcMRI values between each elderly individual and the youth cohort. In the model, the regional fcMRI difference between areas "s" and "i" in the elderly brain relative to areas "i" and "p" in the youth brain is expressed as a connectome distinctiveness index (CDI).
where , is the fcMRI vector between the i-th area and the s-th area in the elderly brain, , is the fcMRI vector between the i-th area and the p-th area in the youth brain, is the number of healthy youth, and , is the fcMRI index between the s-th area and the i-th area in the healthy elderly brain.
The value for each region pair in the elderly brain was estimated relative to that in the youth brain; thus, a total of 98 × 116 matrices were derived for the elderly brain and a total of 90 × 116 matrices for the youth brain.
The distribution of values in the youth brain was evaluated using the equations (2) and Relative to the CDI distribution at the i-th brain area in all youth, the value at the i-th brain area of the s-th elderly individual is defined as This index model can objectively estimate the fcMRI deviation from healthy youth for each AALlabeled brain region, with higher index values 6 indicating a greater degree of deviation in an elderly individual. As 98 healthy elderly individuals were examined, 98 × 116 index matrices were formed.

2.3.3．Identification of functional biomarker regions by fcMRI between the elderly and youth brain
First, the fcMRI indexes for each brain area were averaged across subjects within each age group.
The differences in indexes between the 116 brain regions were calculated using equation (5), and those regions with greatest differences were defined as regions of interest (ROIs).
Between-group differences in fcMRI indexes were evaluated using independent samples t-tests (significance level, α = 5%), and receiver operating characteristic (ROC) curves were constructed using SPSS (IBM SPSS Statistics 21; USA). The BrainNet Viewer toolbox was also employed to highlight functional biomarker regions on a whole-brain template image [24].

2.4.1． Feature vectors
The fcMRI indexes in brain areas identified as functional biomarkers (ROIs) were then considered as feature vectors for the ELM input layer, with those from the poor cognition group labeled as 1 and those from the excellent cognition group labeled as 2.  The accuracy of classification was then evaluated.

2.4.2． ELM model
Among a total of 43 , indexes calculated for the poor cognition group, 34 were randomly selected for the training set and the remaining 9 were used as the testing set. Similarly, of the 55 , indexes in the excellent cognition group, 44 were randomly selected as the training set and the remaining 11 as the testing set. In ELM, the type parameter was set as 1 (Set to 1 to solve the classification problem and set to 0 to solve the regression problem), the number of neurons in the hidden layer was set to 500, and the activation function TF was set to "sig" type. Then, the , indexes were trained and simulated using ELM. Finally, the classification accuracy for the elderly groups was assessed using the testing dataset.

2.4.3． N-fold cross-validation
In ELM, an N-fold cross-validation procedure was employed to test the accuracy of the algorithm as follows. The dataset was divided into parts by setting − 1 parts as the training data and 1 part as the testing data. Thus, classification accuracy could be assessed for each procedure. The mean accuracy over iterations (10 in this case) was utilized to estimate the accuracy of the algorithm.

3.1． FC
The FC matrix diagrams constructed from the rs-fMRI data of elderly individuals with excellent cognitive test scores resembled those constructed from youth brains (Fig. 3).

3.2． Functional biomarkers estimated using the fcMRI index model
To identify regions most sensitive to cognitive test scores, scatter plots of fcMRI index differences between the two groups was constructed (Fig. 4a) Fusiform_R). These five brain areas (Fig. 4b) are therefore potential functional biomarkers (ROIs) for cognitive ability of healthy elderly. Indeed, ROC curves (Fig. 4c) suggested that each regional index distinguished poor from excellent cognitive performance. In addition, p values for all 5 regions were below 5% (Table 1). Thus, FC within subregions of the frontal, parietal, and temporal lobes can distinguish cognitive performance from the fcMRI indexes differences in the healthy elderly's brain.

3.3． Classification using the ELM model
Of the 20 samples in the testing dataset (11 from the excellent cognition group and 9 from the poor group), 18 were correctly classified by the trained ELM model (90% accuracy) (Fig. 5), and the mean accuracy after N-fold cross-validation was 83.5% (Table 2).   5 and Table 2).

4.2． Greater sensitivity of the fcMRI index model compared with conventional FC models
The fcMRI index model developed in this study identified FC changes in single brain areas able to distinguish excellent from poor cognition among healthy elderly individuals with 80%-90% accuracy, whereas conventional FC analysis (Fig. 3)

Data Availability
The data used in this article are from public data sets, and the source of the data is indicated in the article.

Code Availability
The source of the software used in this article has been indicated in the article.

Compliance with Ethical Standards
Conflict of interest All authors declare that they have no conflict of interest.
Ethical Approval This study was approved by the Hebei University of Technology Ethics Committee.

Informed Consent
This study was conducted in accordance with the Declaration of Helsinki. All participants received a thorough explanation of the experimental content in advance and gave their written consent for participation in the study.

Consent for Publication
The Author warrants and represents that the Contribution does not infringe upon any copyright or other rights, and that it does not contain infringing or other unlawful matter, that he/she was the sole and exclusive owner of the rights herein conveyed to the Publisher.